Welcome to my new model page dedicated to the characters from various video games.
Each version will represent a different character from the game, rather than an improvement to the model itself.
I have some ideas regarding future creations, but if you have your propositions - feel free to leave them in the comments and I might make that character! :-)
The base model is a mix of various models that were available in late November (Hassan 1.4 for sure, do not remember others) so the base quality is above pure 1.5.
Currently available:
Street Fighter - Chun-Li
Witcher 3 - Cirilla Fiona Elen Riannon
Cyberpunk 2077 - Female V
Description
This model is quite personal since it is my character of V from Cyberpunk 2077.
This model was done as a case study to see if this idea is viable and it proved to be a total success. I spent with that character around 150 hours and now she comes to life!
NOTE: This is also a nice technique to create consistent characters if you wish to make some comics/graphics/illustrations - just create a character in the game (or some META character) and dreambooth it! As you can see in the examples - it can turn out quite marvelous.
FAQ
Comments (22)
I was a bit confused since your cyberpunk V looks a lot like the militech corpo bitch :D
Meredith Stout, yeah - there is a vibe of her in my V ;-)
Perhaps also why I like Meredith :) But surely Panam and Judy would come first before I even tackle Meredith, and there are so many other nice games to pick from :)
Propositions for new characters:
- Lara Croft / Tomb Raider
- Catwoman from DC
- Batgirl from DC
- Harley Quin from DC
Have you seen these Lara Croft models?
any chance for some imgur albums of them? :)
Character suggestion: Agent 47 (Hitman)
i know this was long time ago, could you prepare some images and put them on imgur? :)
Claire from RE2 Remastered would make my day.
any chance for imgur album of her? :)
I love the reception this receives :-)
My plan is to collect training data on the work days and run the generations on the weekends so expect some nice updates here on saturday/sunday :-)
I think I will make one or two models of my choosing and the rest will be from the comments, so keep them coming! :)
Hi, thank you for your work.
May I ask, Why to train a whole model and not just LoRas, HN or TI? a model size is 2 gb at minimum. HN, TI or LoRas are more suitable if you are going to train characters. Imagine that in a near future you have trained 20 characters. 20 2 gb = 40 GB... In the case of an HN -> 200 mb 20 = 4 GB, a big difference considering space, and the results are almost the same. Maybe a Main model with your style and then HN, TI for the characters... and, if you update your main model, you still can use the HN or TI without a problem.
@nacurutu
hi, thank you for your feedback :-)
the main reason is that i am able to train dreambooth on my machine (using shivamshirao) but i am not able to train hypernetworks nor loras because I am still unable to set it up on my machine, not for the lack of trying :(
i will want to check the TIs and compare to the quality to DB and in the worst case scenario I will be uploading both.
Regarding LORA's - this week has been interesting, there is now a repo that allows us to convert dreambooths to Lora's. I have not checked it yet but I encourage anyone to convert the dreambooths to it and see if they are on the same level. Eventually I will want to configure that myself and upload both versions, for those who value space and those who value quality.
As for the storage issue, I have already trained around 300 models and with the other models I download its around 2 TB of data. It may seem like a lot, but the hard drives are cheap those days :)
That being said, I want to be considerate to those who do not have the ability to store a lot of models. If I could make the Loras/Hypernetworks - I would.
If someone has already managed the DreamBooth -> Lora conversion, please contact me :)
@malcolmrey what i know, LoRa's trainings are on dreambooth also, a model and a LoRa are not the same. a LoRa is the extra information you merge in a regular model training on dreambooth but on a different file. so, I dnt know if a convertion Model -> LoRa could be done.
And, why cant you train HN n TI? Dreambooth needs more hardware(vram) than HN or TI trainings... I can train HN n TI on my 6gb vram card but no Dreambooth training due to my low vram memory amount. Have u tried to train HN n TI using A1111?
@nacurutu because for some reason i can't install the automatic1111 extension for training :/
i use shivamshirao (well, inbd4 fork actually) in WSL2
Love that you're doing this @malcolmrey! I wonder though, if it makes sense to keep these as separate models just to improve discoverability. Certainly each of these deserve to be showcased on their own.
Something that a few community members have requested is the ability to create "packs" of resources. So once we have that functionality, you could still bundle all of your characters together.
Oh, the pack idea would be great. I wonder what other people think about it also.
There were several reasons I decided to do it this way, so I'll do the next model still here to see if it makes sense or not. Perhaps it might be better to split it off. We will see :)
But the reasons were as follows:
* didn't want to have the djz situation, people were not happy that one concept was being spammed all over :)
* there is still this pushback for checkpoint files (i haven't been able to run the A1111 extension, but I've got some tips from scruffy and over the weekend I'll try regular TI so that might change? :P)
* I wanted to use my character as a test case to see how flexible the models are (I knew they could make in-game outputs but wanted to see if real-life output would also be possible, and they are; I got this idea when I managed to make real-life Jinx from Arcane)
* I like the interaction with other users and it seemed like a nice place to gather feedback regarding game characters (instead of spread out over all models :P)
and i did not know about your future feature about packs :)
That all makes sense. I’ll be sure to let you know when the pack stuff is available. Nothing wrong with the way you’re doing it now. Keep doing awesome work 👍
Go deeper, this should likely be 1 model with multi concepts, something almost every dreambooth supports. You don't have to ship one package per truck.
@32bs As far as I can tell so far, there doesn't seem to be a good way to train multi-concept models without a fair amount of bleed between the concepts. The end result is that generally none of the characters look quite like they should.
The hope is that LoRAs will soon be able to be the more portable version of adding data to a checkpoint, but so far they still don't seem to be quite there either.
@justmaier This has been achieved successfully repeatedly in Dungeons and Diffusion, FF7 Diffusion, the endless stream of mixes, etc. StableTuner, naifu, fast-diffusers, ShivamShrirao's, and TheLastBen's forks all support this. OR chain training.
https://huggingface.co/datasets/0xJustin/Dungeons-and-Diffusion
Kinda easy to implement! Get on it civit!
I'd love to see a character model for project eve, now known as stellar blade. Not currently released but some good concept art images around
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.